VLSC Blog

Engineering notes from the VLSC open-source amateur radio projects — remote radio control, SDR signal chains, antenna tuning, and the field work behind them. Written by BG1SB. 中文版

Field · Deep Learning Experiments

Stones from Another Hill: What Deep Learning Actually Did to an FT8 Decoder

On top of a classical chain that already sits at the edge of theory (matched filtering plus LDPC belief propagation), the first question is where the gap is. Measured: the target is the weak-signal tail (69% missed at −26 dB), not same-frequency overlap (the deep configuration already reaches 98.8% of the union). Three attempts then hit three walls — a demapper alone reports 0/30 at −20 dB where the classical chain reports 30/30 (0.937 to the power of 174 is one in ten thousand); both differentiable syndrome forms carry no gradient (1e-19, and exactly zero at high confidence); and a graph-attention decoder trains to the identity while belief propagation sits at 0.0018 on the same soft values. Includes the engineering ledger (data generation was 4.5× the training step) and six questions for people who do this for a living.

BG1SB · 2026-10-15 · about 14 minutes
Field · Fleet Report

Revoked, Still Running: Eleven Days of the Cloud-Hub Era

October 1–11 across the VLSC fleet: ten desktop releases in eleven days and twenty-plus Android releases in eight; a ¥125 W103D box whose MT7663S WiFi was being wedged by the base image's 2020 firmware, cured by the manufacturer's own 2021 N9 (200 packets, 0.5% loss, 6.3 ms, zero drops); spectrum tiers at 151 → 38 → 19 kbps with byte-identical streams for undeclared clients; the FT-891 as eleventh backend; and a revoked Cloud Hub instance that kept running and squatted another callsign's tunnel port. 撤销≠吊销: revocation must stop the thing.

BG1SB · Oct 11, 2026 · ~14 min read
Intelligence · Essay

The Higher the Tune: What the Hands Know and the Gates Keep

Song Yu's two songs are usually read as a ranking. Read against a weekend of byte dumps and firewall rules, they describe two kinds of knowledge instead: one that dies with the session unless converted, one that is worthless until it is. With Zhuangzi's cook and wheelwright, Mencius on the compass and square, and the four times the conversion actually happened.

BG1SB · Oct 3, 2026 · ~13 min read
Field · Cloud Hub

The Borrowed Light: How a Radio Behind NAT Gets an Address

No public IP, no port forwarding, no UPnP — and yet an entry at https://<callsign>.mrrc.vlsc.net/. The four invariants that make it work (outbound only, per-instance certificate pinning, generated routes, a portal with no privileges), the four build-machine defaults that made a running install look dead, and two defects that only exist in a sequence.

BG1SB · Oct 3, 2026 · ~12 min read
Method · Field Report

Reins Side by Side: Three Sessions, Five Days, Twenty-Six Releases

Three model sessions, three repositories, 296 commits, twenty-six releases. The five-day sequel to the forty-eight-hour retrospective: four old root causes returning in new clothes, one new cause surfacing (nothing gates a claim against its own verification), a guard executed by mutation testing, an Android session shipping eight versions in a day and taking three slaps from real hardware, and parallel collaboration going from two collisions to zero. Not by discipline, by structure.

BG1SB · Oct 3, 2026 · five-day edition Oct 5 · ~20 min read
Intelligence · Fourth Position

Pramana: The Fourth Position Above Three Ontologies

Three ontologies are mutually unintelligible, so the way out cannot come from inside any of them. Zhuangzi ruled out all four kinds of arbiter and left one slot open. Something is being built in it now, and it is not a fourth worldview but a warrant protocol drawn on Indian pramana theory: provenance as ring IDs inside attention, warrant as a runtime classifier detecting 91.3% of false absence claims at 15 ms, and standing, which is empty because no actor is required or funded to occupy it. Plus a self-audit: 76 negative assertions across this blog, and not one cell to warrant them.

BG1SB · Oct 1, 2026 · ~30 min read
Intelligence · Field Report

Names and Actualities: What the Field Sent Back

The journal closed on Sep 19 with conclusions grown entirely from one maintainer's practice. Twelve days later a field run came back carrying 2026 measurements. Rectification of names now has a p-value (warrant axis: 1.89 vs 0.95, p = 0.0001); the dregs have been industrialised; flattening is the mathematics of low-rank compression; what got cheap is a counterfeit of judgment; and the outside moved inside the loop. Eight claims from the journal, settled one by one against other people's evidence.

BG1SB · Oct 1, 2026 · ~33 min read
Intelligence · Festival

The Moon Loop: A Mid-Autumn Greeting from a Station That Never Sleeps

Tonight is the 15th of the 8th lunar month: the full moon is both a lantern and radio's oldest reflector. A little moonbounce (echoes since 1953, two-way since 1960, ~2.5 s round trip), tonight's festival-duty roster across MRRC, MRRC Modern, SunMRRC, EFHW and MRRC-FT8, and why the loops keeping watch are exactly what agentic engineering is for. 73 & 88 from BG1SB.

BG1SB · Sep 25, 2026 · ~8 min read
Intelligence

Fate · Faculty · Way · Heaven: The Distillation

The journal's final synthesis in four characters: 命 the terrain you were dealt (the chart, the scorecard), 力 the faculties and their new prices (rent, diagnose, train, earn), 道 the loops that turn energy into learning, 天 the outside that answers. 命曰禀赋,力曰利器,道曰践行,天曰不言 — three cultivable, one only revered: 识命、修力、立道、敬天.

BG1SB · Sep 19, 2026 · ~12 min read
Intelligence · 原文

昆木物语 — 一个理科男的 AI 时代的思考

The day's conversation, preserved verbatim in Chinese — the AI-era reflections of a science guy who scored 149 in physics, 148 in chemistry, full marks in the essay, 128 in Chinese and 87 in math: five numbers, five faculties. From the MRRC × MRRC Modern releases and the FDE board, to 三板斧 and 孤路我独行 and 思接千载, to intent, culture, civilization and the laughter outside the loop, to 斫轮 and 入世/出世 — and to 华盖与伤官. Each message annotated with what it produced.

BG1SB · Sep 19, 2026 · 原文沉淀(中文)
Intelligence

The Faculties: Intellect, IQ, EQ, Knowing, Wisdom, Intuition — and the Two Ways of Living

“Intelligence” is six different things wearing one coat. This philosophical companion to the journal prices each faculty against the AI era — intellect is rentable, IQ devalues as a differentiator, EQ and knowing rise, wisdom and taste can only be earned — and adds the axis that Eastern thought refuses to omit: 入世 / 出世, engagement and transcendence, and why the mature life alternates between them.

BG1SB · Sep 19, 2026 · ~15 min read
Method

Only Imagination Left: When an Idea Becomes an Attempt Immediately

Driving, eating, waking at 3 a.m. — for all of history those sparks had one destination: a notebook, or oblivion. The gap between imagining and trying just collapsed, and that is the largest value this era offers a human being: the daydream has a road. With the honest counterpoint — cheap attempts need governance, evidence, and a verdict that stays human.

BG1SB · Sep 19, 2026 · ~17 min read
Method

The Solo Loop: A One-Person Open-Source Project That Answers Like a Team

No budget, no team, no backing: the nine instruments that let a hobby project give every report a first reading in ten minutes, a checkable answer the same day, and a tested fix in days. What the AI era really changed, the three constraints (attention, continuity, legitimacy), the solo maintainer's operating system, belief with receipts, the failure modes, and a 30-day starting plan.

BG1SB · Sep 19, 2026 · ~20 min read
Method

The Three Axes: From Business Intent to Field Evidence, End to End

MRRC Modern as a worked example of the full lifecycle: one vision sentence → a 15-chapter SDD baseline → 67 version rows in 75 days → field incidents → a support chain that classifies, answers and implements. SDD × Harness × FDE as the method, two threads traced end to end, and eight transferable rules with a three-stage adoption ladder.

BG1SB · Sep 19, 2026 · ~22 min read
Deployment

The Support Loop: How a Week of 🐞 Reports Became Two Releases

Two product lines shipped ten releases in one week — and then the loop itself shipped: a public FDE board where classification is the decision (noise closes, bug auto-schedules, feature waits), and an unattended implementer that lands guarded, test-gated patches on isolated branches while main never moves. The full support chain, the two incidents it absorbed, and what it says about agentic engineering.

BG1SB · Sep 19, 2026 · ~16 min read
Intelligence

Seven Billion Tokens, Dissected: The Ledger That Shrank

The published census of 7,007,437,567 tokens was wrong in three directions at once: overstated by a harness that double-counted chunks of the same API response, understated by 485,213,348 sitting in a table nobody queried, and drifting under log retention and the observer effect. The honest re-run is 7,404,583,808 — and the discipline it produced is snapshot timing, dated evidence, and two ledgers that never vouch for each other.

BG1SB · Sep 5, 2026 · ~18 min read
Intelligence

Connections Rule: Recursive Intelligence from Neurons to Civilization

The argument in four moves: the digital fly shows structure earning testable predictions; connection is skeleton, not organism; the scarce resource is connection quality — six engineering properties, not edge count; and governing connection is the next engineering task.

BG1SB · Aug 28, 2026 · ~26 min read